Donkit Ai

Blog

Notes from the frontier.

Deep dives on agent infrastructure, research, and how we build Donkit.

Why we compile prompts into typed graphs
Engineering
May 12, 2026·8 min

Why we compile prompts into typed graphs

Treating prompts as source code unlocked observability, replays, and safe rollouts.

Read article
Introducing MCP support: bring any tool to Donkit
Product
May 04, 2026·5 min

Introducing MCP support: bring any tool to Donkit

The Model Context Protocol is our universal adapter for the long tail of tools.

Read article
Reducing hallucinations with episodic memory
Research
Apr 21, 2026·11 min

Reducing hallucinations with episodic memory

Episodic recall cut hallucination rates by 38% across our internal benchmarks.

Read article
Cold starts under 300ms on the edge
Engineering
Apr 10, 2026·6 min

Cold starts under 300ms on the edge

How we pre-warm graphs and stream the first token before the runtime is ready.

Read article
Building a 12-person company that ships like 50
Culture
Mar 28, 2026·7 min

Building a 12-person company that ships like 50

Small teams, sharp opinions, brutal focus.

Read article
Donkit 2.0: typed graphs, faster runs, lower bills
Product
Mar 14, 2026·9 min

Donkit 2.0: typed graphs, faster runs, lower bills

Our largest release yet, in one post.

Read article